Add Hermes Agent to local apps

[gary149]

Summary

• Add Hermes Agent as a local app for tool-calling models served via llama.cpp or MLX (same eligibility as Pi: GGUF or MLX, conversational, chat template references tools)
• Snippet uses hermes config set CLI commands (verified to work) instead of editing ~/.hermes/config.yaml by hand
• Refactor: extract isToolCallingLocalAgentModel and getLocalServerStep so Pi and Hermes share the predicate and the server-start step (Pi behavior unchanged byte-for-byte)

Test plan

• pnpm --filter @huggingface/tasks test — 19/19 pass
• Verified end-to-end with unsloth/Qwen3.6-27B-GGUF:UD-Q6_K_XL running on llama.cpp

---

Note

Medium Risk
Medium risk because it changes the generated Pi configuration model.id for GGUF models (now includes the quant tag) and adds new snippet logic that affects what users run/configure, though it’s isolated to local-app guidance.

Overview
Adds Hermes Agent as a new entry in LOCAL_APPS, available for the same tool-calling GGUF/MLX conversational models as Pi, and provides an install/config/run snippet using hermes config set ... against a local OpenAI-compatible server.

Refactors the Pi/Hermes flow by extracting a shared isToolCallingLocalAgentModel predicate and getLocalServerStep snippet generator, and updates Pi’s generated models.json to use the full GGUF model id including :{{QUANT_TAG}} (instead of the repo basename). Tests are expanded to cover Hermes (GGUF + MLX) and updated for the Pi id change.

^{Reviewed by Cursor Bugbot for commit c9c409d. Bugbot is set up for automated code reviews on this repo. Configure here.}

[pcuenca]

🙌

[teknium1]

Looks good from the Hermes side — verified the snippet end-to-end against current main:

• hermes config set model.provider custom / model.base_url / model.default are all valid config paths; custom is a recognized provider and is in the allowlist for vendor-prefixed model IDs, so org/repo:QUANT_TAG won't trigger a doctor warning.
• curl | bash installer URL responds 200, and hermes setup + bare hermes launch flow is correct.
• No changes needed on the hermes-agent side to support this.

Thanks for wiring us up alongside Pi!